[LinalgExt] Better doc for FP8 attention clamping. (iree-org#18301)

Signed-off-by: Stanley Winata <stanley.winata@amd.com>

compiler/src/iree/compiler/Dialect/LinalgExt/Transforms/AggregatedOpInterfaceImpl.cpp

@@ -95,13 +95,15 @@ static Value truncateFloat(OpBuilder &builder, Location loc, AffineMap inputMap,
             APFloat::getLargest(dstTy.getFloatSemantics(), /*Negative=*/false)
                 .convertToDouble();
 
-        // Truncate to the `fp8` range so avoid nan values.
+        // Clamp input to dstTy(usually `fp8`) MAX value to prevent NaNs.
+        // We do not clamp for `-MAX` because this function meant to only be
+        // used by attention's exp2 who's value is always > 0.
         Value mx = builder.create<arith::ConstantOp>(
             loc, builder.getFloatAttr(srcTy, mxDbl));
-        Value sel0 = b.create<arith::MinimumFOp>(loc, mx, args[0]);
+        Value clamped = b.create<arith::MinimumFOp>(loc, mx, args[0]);
 
         // Convert scale to the same datatype as input.
-        Value trunc = convertScalarToDtype(b, loc, sel0, dstTy,
+        Value trunc = convertScalarToDtype(b, loc, clamped, dstTy,
                                            /*isUnsignedCast=*/false);
         b.create<linalg::YieldOp>(loc, trunc);
       });